Microsoft has launched Researcher with Computer Use in Microsoft 365 Copilot , marking a significant advancement in autonomous AI technology. This new capability allows the AI assistant to go beyond simple research tasks and perform actions on behalf of users through a secure virtual computer environment .
This innovation allows Copilot to browse public websites, access authenticated content , and interact with websites while maintaining security standards .
The Computer Use allows the Researcher to access content that requires a subscription and login, perform tasks by clicking buttons and filling out forms, and create complex outputs such as presentations and spreadsheets.
The system operates through a virtual machine running on Windows 365, which acts as a temporary cloud-based computer dedicated to each chat session.
See also: Vulnerability in Claude AI exposes enterprise data
Computer Use changes search capabilities
Users can ask Researcher to prepare customer meeting briefs, gather insights from social media, create personalized reading lists based on current projects, analyze industry trends from limited publications, or turn research findings into well-crafted presentations.
The technology connects to work data, including meetings, files, and conversations, while providing users with full visibility and control.
When enabled, Researcher uses visual browsers, text browsers, terminal interfaces , and the Microsoft Graph to execute integrated workflows.
Users can customize which data sources the system accesses, and enterprise data remains disabled by default when Computer Use is enabled.
The system requests explicit confirmation from the user before performing actions and allows users to take control through secure screen sharing when authentication is required.
Microsoft has implemented strong security measures to address potential risks associated with autonomous AI operations. The virtual machine operates in a fully isolated environment (isolated from corporate networks and user devices).
See also: OpenAI: Aardvark detects errors in code
Safety classifiers inspect every network operation to validate domain security, verify relevance to user queries, and analyze content types. This protection helps attacks cross-prompt injection and jailbreak attempts that can occur while browsing the web.
Browsing actions performed in the sandbox are fully auditable through Microsoft 365 Copilot's logging mechanisms.
User credentials are never transferred to or from the sandbox environment, and all intermediate files are automatically deleted when sessions end.
Administrators control the availability of features through the Microsoft Admin Center, where they can specify which security groups can access Computer Use, manage allow and deny lists, and determine whether users can combine corporate and web data.
See also: Researchers create Linux Rootkit that evades detection
Performance tests show significant improvements, with Researcher achieving 44% better results in BrowseComp benchmarks for complex browsing tasks and a 6% improvement in GAIA assessments.
These benchmarks measure the system's ability to reason across multiple sources of information, synthesize dispersed data, and solve real-world research challenges that require access to diverse datasets and corporate records.
